What is harness engineering?

Harness engineering is the practice of improving an AI agent by improving everything around the model — prompts, tools, skills, evals, memory — using evidence from production. Moda is a harness engineering platform: it maps your harness from the repo, reads production traces, attributes failures to the harness component responsible, and turns them into verified improvements. The model weights stay untouched.

Is Moda a continual learning tool?

Continual learning is the research name for the loop Moda runs: a deployed system that keeps improving from production experience without retraining. Moda's version operates on the agent harness, not the weights — it clusters emergent intents, attributes frustration to root causes, and feeds improvements back into prompts, tools, workflows, retrieval, memory, and evals.

How does Moda improve my agent without retraining the model?

Improvements live in a latent space outside the model weights: prompt edits, tool schema fixes, workflow patches, retrieval tuning, and eval cases derived from real failures. That makes every change portable across models, inspectable, reversible, and per-user adaptive. No fine-tuning pipeline, no weight drift.

What signals does Moda surface from production traces?

Intent discovery via hierarchical clustering of every trace, behavioral failure detection for silent tool misuse and context loss, frustration root cause attribution with agent counterfactuals, and a tool call failure taxonomy. Each signal maps to a concrete harness change, not a dashboard tile.

How is Moda different from a tracing tool like Langfuse or LangSmith?

Tracing tools record what your agent did, span by span. Moda reasons over the whole trace — the full agent run — and across thousands of traces, asking why behavior is changing: which intent clusters are emerging, which tool calls fail in patterns, what root cause explains a frustration spike, and what harness edit closes the loop. Most teams keep their tracer and add Moda for the learning layer.

How does frustration root cause attribution work?

Every frustration event Moda detects ships with a root cause, not just a sentiment score: the trigger turn, the trajectory leading up to it, the user goal that was at risk, an agent counterfactual for what should have happened instead, and the specific harness component (prompt, tool, retrieval step) that caused the failure. Attribution is made over the whole trace, so it reflects what actually happened in the run rather than a single turn.

What LLM providers and frameworks does Moda support?

Provider-agnostic. The Moda SDK supports Node.js and Python applications across major providers and frameworks. Moda also ingests OpenTelemetry and OpenLLMetry traces from OpenAI, Anthropic, Google, OpenRouter, Mistral, Cohere, AWS Bedrock, Azure OpenAI, and any OpenAI-compatible endpoint. Raw JSON ingest is supported for custom agent harnesses.

How long does setup take?

For a straightforward app, a few minutes. Run moda init to connect your workspace and let your coding agent integrate the Moda SDK. Add your API key and ship. Production traces appear in seconds once telemetry is flowing. Hierarchical intent maps are generated during the weekly clustering run once you have at least 50 trace segments.

Harness engineering that developers love

Moda turns production traces into verified improvements for your agent harness: prompts, tools, skills, evals, memory.

Book a demo

LearnedNew signalsWatching

learned · new signals · watching

User correction · "refund still pending"1,284 corrected

Moda learned the canonical policy answer. Prompt patched Mon.

Tool learning · stripe.refund9.6% → 1.4%back to baseline

Schema drift detected. Moda relearned the new arg shape, shipped to 23 agents Tue.

Workflow loop · lookup_order → search_kb38 → 0 sessionseliminated

Loop pattern learned. escalate_to_human gate added, retries dropped 96%.

↗Emerging intent · Apple Pay refunds412 new0 last week

No prior handler. Queued as the next eval + tool.

↗Cohort learning · guest checkout misses218 cohortnew

Moda learned the missing field. Email-only lookup proposed.

⊙Model behavior · professional tone regression+3.1pt refusal vs v3.1monitoring

Behavior change flagged for the next eval set.

Continual learning · Last run 4h ago

Teams shipping with Moda

From broken traces to validated fixes

A continuous system for finding failures, generating improvements & proving what should ship

Book a demo

01 Diagnose production failures

Find where each run broke, why it failed, and whether the issue came from the prompt, tools, workflow, memory, model, or product logic.

02 Generate concrete improvements

Turn repeated failure patterns into specific fixes your team can review: prompt changes, tool updates, workflow edits, verifier gates, eval cases, and reusable skills.

03 Validate what to ship

Test improvements against historical production traces, measure impact and regressions, then recommend the most reliable fix.

Find out where each run broke and why

Moda analyzes agent logs to separate prompt, tool, workflow, memory, model, and product-logic failures — across the whole trace.

Production traces

Failure diagnosis

Runs

1.2k

sampled today

Failed

14%

clustered

Cause

families

Refund flow stalled

tool returned stale policy version

tool

Memory note reused

old user preference overrode session

memory

Handoff gap

agent skipped escalation threshold

workflow

Prompt ambiguity

two intents mapped to same action

prompt

Root cause

stale policy lookup

first seen14:07

affected flowsrefund, policy, billing

confidence92%

tool mismatch42%

prompt drift21%

workflow loop13%

Know what will improve before you ship

Replay candidate improvements against historical production traces, measure impact and regressions, then ship with evidence.

Replay evals

Validated ship decision

Candidatequalitycostrisk

baseline

72%$0.19med

prompt guard

81%$0.18low

tool adapter + gate

91%$0.16low

workflow rewrite

88%$0.24high

model swap

85%$0.21med

Applied to winner

tool adapternormalize refund reason

verifier gateblock stale policy

eval casereplay 86 refund traces

Recommended

+19%

quality lift, safe to merge

Regression check

pass

refund traces86/86

policy citations79/82

handoff latency-24%

eval suite128/130

cost / run-12%

Ship decision

tool adapter + gate

ready

+19% quality · 0 regressionsmerge ready

Repeated failures become reviewable changes

Moda turns recurring patterns into specific prompt, tool, workflow, verifier, eval, and skill improvements your team can inspect.

Trace-to-fix

Generated improvements

Patterns

ranked by impact

tool schema drift

stripe.refund requires reason

fix

missing verifier

policy answer lacks citation

gate

loop detector

lookup_order repeats twice

skill

generate patch: prompt guard, tool adapter, eval case, reviewer checklist

Patch plan

Prompt guard

Require source-backed policy answer before final response

ready

Tool adapter

Normalize refund reason and retry failed schema calls

ready

Eval cases

Replay 86 refund traces with stale-policy variants

new

Verifier gate

Block answer when policy version is older than session

new

Built around the
trace to fix loop

Diagnose where agent runs break, generate concrete harness improvements, and validate candidate fixes against historical traces before your team ships.

Book a demo

Diagnose

Failure sources mapped: 94%
Median time to root cause: 6 min
Traces analyzed weekly: 1.2M

Failure-source coverage · higher is better

Frequently asked questions

Everything you need to know before connecting your first production traces.

Learn more

Harness engineering that developers love

From broken traces to validated fixes

01 Diagnose production failures

02 Generate concrete improvements

03 Validate what to ship

Find out where each run broke and why

Know what will improve before you ship

Repeated failures become reviewable changes

Built around the trace to fix loop

Frequently asked questions

What is harness engineering?

Is Moda a continual learning tool?

How does Moda improve my agent without retraining the model?

What signals does Moda surface from production traces?

How is Moda different from a tracing tool like Langfuse or LangSmith?

How does frustration root cause attribution work?

What LLM providers and frameworks does Moda support?

How long does setup take?

Built around the
trace to fix loop